Re: [Qemu-devel] [PATCH] 1/4] target-ppc: Implement bcdcfn. instruction

[joserz@linux.vnet.ibm.com]

Hello David,

Thank you very much for your review. As you might have noticed this is my first patch so I hope you don't mind about my newbie questions/explanations below.

On Thu, Oct 27, 2016 at 12:05:30PM +1100, David Gibson wrote:
> On Wed, Oct 26, 2016 at 11:18:55AM -0200, Jose Ricardo Ziviani wrote:
> > bcdcfn. converts from National numeric format to BCD. National format
> > uses a byte to represent a digit where the most significant nibble is
> > always 0x3 and the least sign. nibbles is the digit itself.
> > 
> > Signed-off-by: Jose Ricardo Ziviani <joserz@linux.vnet.ibm.com>
> > ---
> >  target-ppc/helper.h                 |  1 +
> >  target-ppc/int_helper.c             | 54 ++++++++++++++++++++++++++
> >  target-ppc/translate/vmx-impl.inc.c | 75 +++++++++++++++++++++++++++++++++++++
> >  target-ppc/translate/vmx-ops.inc.c  |  4 +-
> >  4 files changed, 132 insertions(+), 2 deletions(-)
> > 
> > diff --git a/target-ppc/helper.h b/target-ppc/helper.h
> > index 04c6421..d30ec60 100644
> > --- a/target-ppc/helper.h
> > +++ b/target-ppc/helper.h
> > @@ -369,6 +369,7 @@ DEF_HELPER_4(vpermxor, void, avr, avr, avr, avr)
> >  
> >  DEF_HELPER_4(bcdadd, i32, avr, avr, avr, i32)
> >  DEF_HELPER_4(bcdsub, i32, avr, avr, avr, i32)
> > +DEF_HELPER_3(bcdcfn, i32, avr, avr, i32)
> >  
> >  DEF_HELPER_2(xsadddp, void, env, i32)
> >  DEF_HELPER_2(xssubdp, void, env, i32)
> > diff --git a/target-ppc/int_helper.c b/target-ppc/int_helper.c
> > index 5aee0a8..494c74e 100644
> > --- a/target-ppc/int_helper.c
> > +++ b/target-ppc/int_helper.c
> > @@ -2417,6 +2417,8 @@ void helper_vsubecuq(ppc_avr_t *r, ppc_avr_t *a, ppc_avr_t *b, ppc_avr_t *c)
> >  #define BCD_NEG_PREF    0xD
> >  #define BCD_NEG_ALT     0xB
> >  #define BCD_PLUS_ALT_2  0xE
> > +#define NATIONAL_PLUS   0x2B
> > +#define NATIONAL_NEG    0x2D
> >  
> >  #if defined(HOST_WORDS_BIGENDIAN)
> >  #define BCD_DIG_BYTE(n) (15 - (n/2))
> > @@ -2483,6 +2485,15 @@ static void bcd_put_digit(ppc_avr_t *bcd, uint8_t digit, int n)
> >      }
> >  }
> >  
> > +static uint8_t get_national_digit(ppc_avr_t *reg, int n)
> > +{
> > +#if defined(HOST_WORDS_BIGENDIAN)
> > +    return reg->u16[8 - n] & 0xFF;
> > +#else
> > +    return reg->u16[n] & 0xFF;
> > +#endif
> 
> You're discarding the high byte of each digit here, which means you
> won't detect badly encoded values that have correct low bytes.

OK! The high byte is expected to be 0, I'll check if it's != 0 and set a
flag like:

*invalid = (reg->u16[n] >> 8) != 0;

makes sense?

> 
> > +}
> > +
> >  static int bcd_cmp_mag(ppc_avr_t *a, ppc_avr_t *b)
> >  {
> >      int i;
> > @@ -2613,6 +2624,49 @@ uint32_t helper_bcdsub(ppc_avr_t *r,  ppc_avr_t *a, ppc_avr_t *b, uint32_t ps)
> >      return helper_bcdadd(r, a, &bcopy, ps);
> >  }
> >  
> > +uint32_t helper_bcdcfn(ppc_avr_t *r, ppc_avr_t *b, uint32_t ps)
> > +{
> > +    int i;
> > +    int is_zero = 0;
> > +    int cr = 0;
> > +    int national = 0;
> > +    ppc_avr_t ret = { .u64 = { 0, 0 } };
> > +    uint16_t sgnb = get_national_digit(b, 0);
> > +    int invalid = (sgnb != NATIONAL_PLUS && sgnb != NATIONAL_NEG);
> > +
> > +    for (i = 1; i < 8; i++) {
> > +        national = get_national_digit(b, i);
> > +
> > +        is_zero += (national == 0x30);
> > +        if (unlikely(national < 0x30 || national > 0x39)) {
> > +            invalid = 1;
> > +        }
> > +
> > +        bcd_put_digit(&ret, national & 0xf, i);
> > +    }
> > +
> > +    if (sgnb == NATIONAL_PLUS ||
> > +            (b->u64[0] == 0 && b->u64[1] == 0)) {
> 
> The second part of this test doesn't seem to make much sense.  I think
> you're trying to always put a positive sign on a zero result.  But
> you're testing the national encoded input, not the BCD encoded output,
> and zero will *not* be all zero bits in national encoding.

OK! ISA defines invalid encoding as undefined behavior. I forced a
positive sign to vrt when vrb = 0 (invalid enc.) but it is not necessary.
I'll remove it in v2.

> 
> > +        bcd_put_digit(&ret, (ps == 0) ? BCD_PLUS_PREF_1 : BCD_PLUS_PREF_2, 0);
> > +    } else {
> > +        bcd_put_digit(&ret, BCD_NEG_PREF, 0);
> > +    }
> > +
> > +    if (!is_zero) {
> 
> From the logic above, 'is_zero' is currently a count of how many 0
> digits the value has, not whether the value as a whole is zero.  That
> means you will get the wrong result here.
> 

OK! I'll fix that logic.

> > +        cr = (sgnb == NATIONAL_PLUS) ? 1 << CRF_GT : 1 << CRF_LT;
> > +    } else {
> > +        cr = 1 << CRF_EQ;
> > +    }
> > +
> > +    if (unlikely(invalid)) {
> > +        cr = 1 << CRF_SO;
> > +    }
> > +
> > +    *r = ret;
> > +
> > +    return cr;
> > +}
> > +
> >  void helper_vsbox(ppc_avr_t *r, ppc_avr_t *a)
> >  {
> >      int i;
> > diff --git a/target-ppc/translate/vmx-impl.inc.c b/target-ppc/translate/vmx-impl.inc.c
> > index c8998f3..2abdcac 100644
> > --- a/target-ppc/translate/vmx-impl.inc.c
> > +++ b/target-ppc/translate/vmx-impl.inc.c
> > @@ -871,8 +871,81 @@ static void gen_##op(DisasContext *ctx)             \
> >      tcg_temp_free_i32(ps);                          \
> >  }
> >  
> > +#define GEN_BCD2(op)                                \
> > +static void gen_##op(DisasContext *ctx)             \
> > +{                                                   \
> > +    TCGv_ptr rd, rb;                                \
> > +    TCGv_i32 ps;                                    \
> > +                                                    \
> > +    if (unlikely(!ctx->altivec_enabled)) {          \
> > +        gen_exception(ctx, POWERPC_EXCP_VPU);       \
> > +        return;                                     \
> > +    }                                               \
> > +                                                    \
> > +    rb = gen_avr_ptr(rB(ctx->opcode));              \
> > +    rd = gen_avr_ptr(rD(ctx->opcode));              \
> > +                                                    \
> > +    ps = tcg_const_i32((ctx->opcode & 0x200) != 0); \
> > +                                                    \
> > +    gen_helper_##op(cpu_crf[6], rd, rb, ps);        \
> > +                                                    \
> > +    tcg_temp_free_ptr(rb);                          \
> > +    tcg_temp_free_ptr(rd);                          \
> > +    tcg_temp_free_i32(ps);                          \
> > +}
> > +
> >  GEN_BCD(bcdadd)
> >  GEN_BCD(bcdsub)
> > +GEN_BCD2(bcdcfn)
> > +
> > +static void gen_xpnd04_1(DisasContext *ctx)
> > +{
> > +    switch (opc4(ctx->opcode)) {
> > +    case 0:
> > +        break; /* bcdctsq. */
> > +    case 2:
> > +        break; /* bcdcfsq. */
> > +    case 4:
> > +        break; /* bcdctz. */
> > +    case 5:
> > +        break; /* bcdctn. */
> > +    case 6:
> > +        break; /* bcdcfz. */
> > +    case 7:
> > +        gen_bcdcfn(ctx);
> > +        break;
> > +    case 31:
> > +        break; /* bcdsetsgn. */
> > +    default:
> > +        break;
> > +    }
> > +}
> > +
> > +static void gen_xpnd04_2(DisasContext *ctx)
> > +{
> > +    switch (opc4(ctx->opcode)) {
> > +    case 0:
> > +        break; /* bcdctsq. */
> > +    case 2:
> > +        break; /* bcdcfsq. */
> > +    case 4:
> > +        break; /* bcdctz. */
> > +    case 6:
> > +        break; /* bcdcfz. */
> > +    case 7:
> > +        gen_bcdcfn(ctx);
> > +        break;
> > +    case 31:
> > +        break; /* bcdsetsgn. */
> > +    default:
> > +        break;
> > +    }
> > +}
> 
> I thougt the main opcode table now had support for opc4, so you
> shouldn't need these special expanders.

This is the hardest one. :)

For all the functions above we have opcode collisions today. Let's take the bcdcfn case, its opcodes are:

000100 vrt:5 00111 vrb:5 11110 00000 1, when ps=1
000100 vrt:5 00111 vrb:5 10110 00000 1, when ps=0

Registering the opcode in the opcode table at translate_init.c, the following sequence is assumed:

0x4 -> 0x0 -> 0x1e (ps=1) or 0x16 (ps=0) -> 0x7

However,

0x4, 0x0, 0x1e already maps to a direct opcode:
vsubsws 000100 vrt:5 vra:5 vrb:5 11110 00000 0

as well as  0x4, 0x0, 0x16, that maps to:
vsubcuw 000100 vrt:5 vra:5 vrb:5 10110 00000 0

I this case, gen_xpnd04_1 handles ps=0 collisions and gen_xpnd04_2 does the same when ps=1. So Nikunj gave me this great idea to handle these exceptional cases.

> 
> > +GEN_VXFORM_DUAL(vsubcuw, PPC_ALTIVEC, PPC_NONE, \
> > +                xpnd04_1, PPC_NONE, PPC2_ISA300)
> > +GEN_VXFORM_DUAL(vsubsws, PPC_ALTIVEC, PPC_NONE, \
> > +                xpnd04_2, PPC_NONE, PPC2_ISA300)
> >  
> >  GEN_VXFORM_DUAL(vsububm, PPC_ALTIVEC, PPC_NONE, \
> >                  bcdadd, PPC_NONE, PPC2_ALTIVEC_207)
> > @@ -949,3 +1022,5 @@ GEN_VXFORM_DUAL(vsldoi, PPC_ALTIVEC, PPC_NONE,
> >  #undef GEN_VXFORM_NOA
> >  #undef GEN_VXFORM_UIMM
> >  #undef GEN_VAFORM_PAIRED
> > +
> > +#undef GEN_BCD2
> > diff --git a/target-ppc/translate/vmx-ops.inc.c b/target-ppc/translate/vmx-ops.inc.c
> > index 68cba3e..7a5fec6 100644
> > --- a/target-ppc/translate/vmx-ops.inc.c
> > +++ b/target-ppc/translate/vmx-ops.inc.c
> > @@ -122,7 +122,7 @@ GEN_VXFORM_300(vslv, 2, 29),
> >  GEN_VXFORM(vslo, 6, 16),
> >  GEN_VXFORM(vsro, 6, 17),
> >  GEN_VXFORM(vaddcuw, 0, 6),
> > -GEN_VXFORM(vsubcuw, 0, 22),
> > +GEN_VXFORM_DUAL(vsubcuw, xpnd04_1, 0, 22, PPC_ALTIVEC, PPC_NONE),
> >  GEN_VXFORM(vaddubs, 0, 8),
> >  GEN_VXFORM(vadduhs, 0, 9),
> >  GEN_VXFORM(vadduws, 0, 10),
> > @@ -134,7 +134,7 @@ GEN_VXFORM_DUAL(vsubuhs, bcdsub, 0, 25, PPC_ALTIVEC, PPC_NONE),
> >  GEN_VXFORM(vsubuws, 0, 26),
> >  GEN_VXFORM(vsubsbs, 0, 28),
> >  GEN_VXFORM(vsubshs, 0, 29),
> > -GEN_VXFORM(vsubsws, 0, 30),
> > +GEN_VXFORM_DUAL(vsubsws, xpnd04_2, 0, 30, PPC_ALTIVEC, PPC_NONE),
> >  GEN_VXFORM_207(vadduqm, 0, 4),
> >  GEN_VXFORM_207(vaddcuq, 0, 5),
> >  GEN_VXFORM_DUAL(vaddeuqm, vaddecuq, 30, 0xFF, PPC_NONE, PPC2_ALTIVEC_207),
> 
> -- 
> David Gibson			| I'll have my music baroque, and my code
> david AT gibson.dropbear.id.au	| minimalist, thank you.  NOT _the_ _other_
> 				| _way_ _around_!
> http://www.ozlabs.org/~dgibson